{"id":"3cff4ac8-659c-4afb-8a6e-84b61beb7003","arxiv_id":"2501.17968","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 100-ms online trajectory replanner, combined with a 10-second offline plan, lets a robot dynamically grasp irregular objects despite pose estimation errors.","lead":"This paper describes a two-stage trajectory planner that lets a robot arm grasp irregularly shaped objects while the arm keeps moving. It first computes a full path offline in about 10 seconds, then recomputes small corrections online every 100 milliseconds whenever the camera updates the object's position.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The small-deviation linearization in Eqs. (10)-(11) is untested at the pose-error magnitudes the method is meant to handle; one real miss and no sensitivity sweep leave the central robustness claim unsecured.","rationale":"The reader's weakest assumption points to the same load-bearing step: the linearization in Eqs. (10)-(11). My attack sharpens it: because the state-control dynamics f(x,v)=[qdot; v] is affine in x and v, the first-order expansion is exact for state/control deviations alone; the only dropped term is the product of the time perturbation with the linearized state deviation. Therefore the real question is the joint size of (deltaT, delta x, delta v) relative to the curvature of the nominal trajectory, not simply the nonlinearity of f. The bounds in (13c) are constant boxes and do not certify that the linearized equality constraints approximate the true trapezoidal integration within an acceptable tolerance. This matters because the abstract and introduction explicitly motivate the replanner as a response to pose-estimation errors; if those errors are large enough, the online step can fail or return a trajectory that violates the true dynamics. The paper's 20-trial moving-object experiment includes one miss and no error-magnitude sweep, so the evidence does not yet bound the regime of validity. I agree with the conditional verdict: the architecture, the 10 Hz/100 ms timing, and the real demonstrations are genuinely useful, and the proposed test would either confirm the linearization's practical range or reveal a concrete failure mode. No change to the verdict is needed beyond the reader's already-requested revisions.","tokens_in":10603,"tokens_out":8192,"duration_ms":79370,"concrete_test":"Re-run the 1000-pose simulation with injected pose-estimation errors of increasing magnitude (translation 0, 2, 5, 10, 20 mm; rotation 0, 1, 2, 5, 10 deg) around the nominal object pose. For each level, record: (a) feasibility and solve time of the LCQP (13); (b) the residual of the true nonlinear discrete-time rollout x_{k+1}=x_k+DeltaT/(2N)(f_k+f_{k+1}) evaluated on the replanned trajectory (15); (c) grasp success. If success drops or the rollout residual grows sharply before reaching the vision system's expected error bound, the small-deviation assumption is the limiting factor; if success remains high throughout, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the online replanner (13) reliably compensates pose-estimation errors while the robot stays in motion. That claim rests on the first-order expansion (10)-(11), where products of deviation variables are dropped and the time perturbation enters via the nominal increment (x*_{k+1}-x*_k). Although f(x,v)=[qdot; v] is affine, the neglected term is the joint time-state perturbation deltaT * Gamma * delta, so the approximation is valid only when |deltaT| and ||delta x|| are small relative to the nominal trajectory. The box bounds (13c) do not encode any validity radius for this linearization, and the QP has no recovery mechanism if it is infeasible or hits the iteration limit (as happened in 12/1000 simulation trials). If a vision update or sudden object relocation moves the target beyond that radius, the replanned trajectory can fail to satisfy the true discrete-time dynamics or miss the updated grasp pose. The moving-object evidence is 20 trials with 1 miss; no experiment varies pose-error magnitude. Thus the paper's motivating robustness claim is unsupported exactly where the method's novelty lies. The authors' self-identified limitation in Section V is the 10 s offline compute; that is real but distinct from the untested linearization regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage trajectory optimization framework for dynamically grasping irregular objects with a KUKA LBR iiwa and SDH2 gripper. An offline stage computes three-phase trajectories (approach, grasp, transfer) in under 10 s, and an online replanner solves a small quadratic program at 10 Hz to adjust the trajectory when the vision system updates the object pose. The authors report a 98.8% success rate over 1000 simulated random object poses, an average online solve time of 53.5 ± 7.9 ms, and real-robot experiments including a single-trial with/without comparison and a 20-trial moving-object comparison against RRTConnect (95% vs. 75% success). The central claim is that the online replanner makes dynamic grasping robust to pose-estimation errors while the robot remains in motion.","tokens_in":10808,"tokens_out":4798,"duration_ms":51457,"significance":"If the results hold, the paper offers a practically fast replanning scheme with a clear timing budget, a principled offline warm-start, and an interesting two-stage use of trajectory optimization. The strengths are explicit: offline computation is measured against VP-STO, online computation is well below the 100 ms budget, the simulation campaign is large (1000 trials), and the method is demonstrated on real hardware with two different grippers. The derivation of the online QP is internally consistent, and the use of torque bounds via previous work is a reasonable modeling choice. However, the evidence for the central robustness claim is not yet conclusive: the linearization in the online replanner is tested only indirectly, the key with/without online comparison has n=1 per condition, and the moving-object comparison is a 20-trial experiment reported without confidence intervals. These gaps are fixable and do not, in my reading, invalidate the proposed approach.","major_comments":[{"comment":"The linearized constraint (11) drops products of δT with δx_k and δv_k, so it is reliable only when deviations from the nominal trajectory are small. However, the box constraints in (13c) do not encode any validity radius for this linearization, and no experiment or simulation varies the magnitude of the pose-estimation error. I request a sensitivity study: apply synthetic pose offsets of increasing magnitude at representative object poses, solve (13), and report the true trapezoidal residual of the replanned trajectory, the QP solve success rate, and the resulting grasp success rate. Without such a study, the paper's motivating claim that the replanner handles vision pose errors is unverified precisely in the regime where the dropped product terms become large.","section":"III-B, Eqs. (10)-(13)"},{"comment":"The comparison of the proposed framework with and without the online replanner is based on one trial per condition. The conclusion that the replanner is responsible for the successful grasp cannot be drawn from n=1; the two trajectories in Fig. 7 and snapshots in Fig. 8 are illustrative only. I recommend repeating the experiment over at least 10-20 random object placements with the same initial and target configurations and reporting success rates, per-trial deviation magnitudes, and, where possible, grasp error or minimum clearance.","section":"IV-B, Fig. 7-8"},{"comment":"The 20-trial moving-object comparison reports 95% versus 75% success without confidence intervals or a statistical test. With 19/20 successes, the one-sided 95% Wilson lower bound is approximately 76%, so the data do not strongly separate the proposed method from RRTConnect, especially since two RRTConnect trials violated the time limit, which may reflect the TOTG time parametrization rather than planning quality. Please report exact per-trial outcomes, confidence intervals, and either a paired test or a larger trial count.","section":"IV-B, Table II"},{"comment":"The paper states that in 12 failed simulation trials 'the optimization problems reach the iteration limits set to 100,' but it is not specified whether these failures occur in the offline phase, the online QP, or both, and no recovery mechanism is described. Because the 98.8% success rate is a headline quantitative result, the failure stage and subsequent behavior matter. Please clarify where the failures occur, report the distribution of the resulting deviations, and discuss any fallback or re-initialization if the online QP becomes infeasible or hits its iteration limit.","section":"IV-A (Simulation results)"}],"minor_comments":[{"comment":"The caption says 'RRTConnect [25]', but reference [25] is the MoveIt paper while the RRT-Connect algorithm is reference [26]; please correct the citation.","section":"Table II"},{"comment":"The inequality (13c) uses the same symbol s_k for the optimization variable and its lower bound; please introduce distinct notation such as \\underline{s}_k and \\overline{s}_k to make the box constraints unambiguous.","section":"Eq. (13c)"},{"comment":"The numerical values of the weighting matrices Q_k in Eq. (14), the admissible bound vectors in (13c), and the potential-function parameters (z_th, μ, Σ) in Eq. (5) are not reported; providing these values would improve reproducibility.","section":"III-B, Eq. (14), and III-A"},{"comment":"The caption states σ_x² = 5 mm and σ_y² = 2 mm, which are variances with units of mm²; please clarify whether these are variances or standard deviations and use consistent units.","section":"Fig. 5"},{"comment":"The text says 'AruCo marker'; the standard spelling is 'ArUco marker'.","section":"I.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent application paper, but in its current form the evidence for the load-bearing robustness claim is narrower than the abstract implies. The central algorithm is plausible and the experiment count is reasonable for an initial study, yet the absence of a pose-error sensitivity analysis and the small sample of the key comparison are the main barriers to acceptance. These concerns are addressable within the scope of the manuscript, so I recommend major revision rather than rejection. I have no concerns about citation fairness or duplicate submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid engineering contribution to reactive grasping, and the 100 ms online replanner is a genuine result, but the robustness story is not as strong as the titles and claims imply. The statistical evidence is thin in exactly the places that matter.\n\nWhat's new: the paper integrates a three-phase offline trajectory optimization (approach, grasp, deliver) with a lightweight online replanner that linearizes the dynamics around the offline trajectory and solves a small QP in about 53 ms. That timing is credible and below the 100 ms budget. The simulation over 1000 random poses (988 successes) and the real 20-trial moving-object experiment (19/20) are reasonable demonstrations for a robotics paper. The writing is clear, the math is internally consistent, and the authors are honest about the 10 s offline compute limitation.\n\nWhere the soft spots are: the with/without online replanner comparison is a single trial each, so it's anecdotal, not evidence. The moving-object comparison against RRTConnect is 20 trials with no confidence intervals, and RRTConnect is not a fair reactive baseline because it is a geometric planner with time parametrization bolted on. The computation times exclude failed trials, which biases the timing statistics. No code or parameter values are shipped, so the results are not reproducible as-is.\n\nThe deeper concern, which the stress-test correctly identifies, is the linearization in Eqs. (10)-(11). Dropping products of deviations is fine when the pose error is small, but the paper never shows how large an error the replanner can tolerate. The QP has bounds but no validity radius, and 12/1000 simulation trials hit the iteration limit. That said, the real experiments suggest the method does handle the errors that actually occurred in their setup. So I'd call this a missing sensitivity analysis rather than a fatal flaw. The authors should add a sweep over pose-error magnitude and report successes and failures as a function of error size.\n\nWho it's for: readers working on reactive manipulation, dynamic grasping, or fast trajectory optimization. It's a useful recipe with honest limitations, but it is not a theoretical breakthrough.\n\nRecommendation: send it to peer review. A serious referee would ask for the sensitivity analysis and better statistics, but the core idea and the real-robot evidence merit a revision, not a desk reject.","headline":"Solid engineering paper on reactive trajectory replanning for dynamic grasping, but the robustness claims rest on thin statistics and an untested linearization regime.","tokens_in":11384,"tokens_out":3442,"would_cite":false,"duration_ms":35293,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage trajectory replanner lets robots dynamically grasp irregular objects despite vision pose errors.","keywords":["dynamic grasping","trajectory replanning","irregular object manipulation","online trajectory optimization","linear constrained quadratic program","pose estimation error compensation","moving object grasping","robot manipulation"],"falsifier":"Run the online replanner on the same irregular-object grasp while injecting controlled position biases into the vision module (for example, 1, 5, 10, and 20 mm) and record both grasp success and QP solver feasibility; a sharp drop in success or repeated solver failures at moderate bias would show that the linearization, rather than the tracking controller, is the limiting factor.","tokens_in":10334,"feed_emoji":"🤖","tokens_out":7927,"duration_ms":74878,"temperature":0.7,"pith_summary":"The paper aims to show that dynamic grasping—catching or picking an object while the robot keeps moving—can work for irregular, non-symmetric objects, not just blocks or boxes. The proposed framework splits planning into an offline stage that produces three smooth motion segments (approach, grasp, carry-away) in under 10 seconds, and an online replanner that reacts to camera pose updates at 10 Hz by solving a small quadratic program in about 53 milliseconds. The online update is built by linearizing the robot dynamics around the offline trajectory, so it nudges the arm and gripper to the latest object pose without stopping the motion. According to the paper, 988 of 1000 randomized simulated poses were grasped successfully, and in real moving-object trials the replanner caught the object 19 out of 20 times, compared with 15 out of 20 for an RRTConnect baseline. If this holds up, the method would make production lines more flexible by letting robots grasp randomly placed or moving irregular parts while in motion.","feed_headline":"Online replanner catches moving irregular objects 19 of 20 times","feed_subtitle":"A 53 ms linearized update absorbs vision noise, lifting success from 75% to 95% on real moving objects.","key_machinery":"The load-bearing object is the deviation trajectory $\\delta\\xi = [\\delta T, \\delta x_0^T, \\ldots, \\delta x_N^T, \\delta v_0^T, \\ldots, \\delta v_N^T]$ and the linearized discrete-time map that relates deviations at neighboring grid points. Starting from the offline optimal trajectory $\\xi^*$, the replanner computes Jacobians $\\Gamma^x_k$ and $\\Gamma^v_k$ of the dynamics, neglects second-order terms, and obtains the compact linear map $C_{k+1}s_{k+1}=A_k s_k$, where $s_k=[\\delta x_k^T,\\delta v_k^T,\\delta T_k]^T$. The quadratic program (13) then minimizes a weighted sum of squared deviations subject to that map, box constraints on deviations, and boundary conditions set by the updated object pose. In the offline phase, an inverted-Gaussian potential function restricts the gripper's contact point inside a funnel near the object, and torque limits are respected by bounding the Coriolis term rather than neglecting it.","core_discovery":"The paper's central claim is that the combination of a fast offline trajectory optimizer and an online deviation-minimizing replanner makes dynamic grasping of irregular objects robust to pose-estimation error. The offline stage solves a constrained trajectory optimization for three connected phases, with the final state of each phase serving as the initial state of the next, and uses an inverted-Gaussian potential function to keep the gripper approaching the object along a safe funnel. The online stage updates the whole trajectory whenever a new object pose arrives: it introduces a small deviation vector $\\delta\\xi$ around the current optimal trajectory, linearizes the discretized dynamics at the reference trajectory, drops products of deviation variables, and solves a linear constrained quadratic program (Eq. 13) whose solution is added back to the reference. Reported running times are below 10 s for all three offline phases and $53.5 \\pm 7.9$ ms on average for the online replanner; success rates are 988/1000 in simulation and 19/20 for real moving objects, versus 15/20 for RRTConnect.","pith_inferences":["The linearization assumption implies there is a basin of attraction: success should degrade continuously as pose-estimation bias grows, so a natural extension not tested in the paper is a sensitivity sweep over injected vision errors to find the breakpoint.","The same online-replanner structure could apply beyond grasping, e.g., to assembly insertion or handover tasks where a goal pose stream arrives at low rate and the robot must correct while moving.","Because the replanner only produces local corrections around the offline trajectory, completely unexpected object relocations may still require a global planner; the generalization experiment only covers one sudden relocation.","The authors' stated limitation (the 10 s offline computation) suggests a fully online version with shorter planning horizons is the immediate next step; if learned motion primitives replace the offline stage, the two-phase split could disappear entirely."],"forward_implications":["A robot can grasp randomly placed irregular parts without stopping: the online replanner absorbs 10 Hz pose updates within a 100 ms budget, so the arm can keep moving through approach and grasp.","Dynamic grasping success in the paper's tests is 98.8% in 1000 randomized simulated poses and 95% over 20 real moving-object trials.","Compared with an RRTConnect baseline, the replanner avoids computation-time limit violations (0 versus 2 trials) and misses the object less often (1 versus 3 trials).","The same two-stage framework transfers to a different gripper, indicating the method is not tied to one hand's kinematics.","Because the online update is a convex quadratic program, it can run on a standard PC, making the approach practical for industrial controllers."],"supporting_citations":[{"why":"Supplies the fast offline trajectory optimization technique and the way torque limits are handled by bounding the Coriolis matrix, which the present framework builds on.","marker":"[10]"},{"why":"Provides the via-point stochastic trajectory optimization baseline whose offline computation times are compared in Table I.","marker":"[13]"},{"why":"IPOPT, the interior-point solver used to solve both the offline trajectory optimization (3) and the online replanner's linear constrained quadratic program (13).","marker":"[22]"},{"why":"MA57, the sparse linear solver called by IPOPT in the reported computation times.","marker":"[24]"},{"why":"Supplies MoveIt, the software environment in which the RRTConnect baseline is run.","marker":"[25]"},{"why":"Defines the RRTConnect single-query planning algorithm used as the baseline in the real moving-object comparison.","marker":"[26]"},{"why":"Provides the time-optimal trajectory generation used to give the RRTConnect path a time parametrization.","marker":"[27]"}],"fun_headline_variants":["53 ms replan boosts dynamic grasp success to 95% in real time","Online replanner lifts grasping robustness to 95% from 75%","98.8% sim, 95% real: online replanner for dynamic grasps","Fast replanner tames pose error for irregular-object grasps","Replanner achieves 95% success on real dynamic grasps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pose errors and disturbances are small enough that the linearized dynamics with products of deviation variables dropped remain valid; if a vision error or sudden object relocation pushes the system out of that small-deviation regime, the quadratic program can fail or steer the gripper past the object.","fun_headline_variants_meta":{"raw":{"variants":["53 ms replan boosts dynamic grasp success to 95% in real time","Online replanner lifts grasping robustness to 95% from 75%","98.8% sim, 95% real: online replanner for dynamic grasps","Fast replanner tames pose error for irregular-object grasps","Replanner achieves 95% success on real dynamic grasps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001084,"raw_usage":{"total_tokens":4530,"prompt_tokens":938,"completion_tokens":3592,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":3494}},"tokens_in":554,"tokens_out":3592,"duration_ms":22600,"temperature":1.0,"reasoning_tokens":3494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:27:37.127099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the online replanner on the same irregular-object grasp while injecting controlled position biases into the vision module (for example, 1, 5, 10, and 20 mm) and record both grasp success and QP solver feasibility; a sharp drop in success or repeated solver failures at moderate bias would show that the linearization, rather than the tracking controller, is the limiting factor.","supporting_citations":[{"cited_title":"Fast swing-up trajectory optimization for a spherical pendulum on a 7-dof collaborative robot,","cited_arxiv_id":null,"evidence_quote":"Supplies the fast offline trajectory optimization technique and the way torque limits are handled by bounding the Coriolis matrix, which the present framework builds on."},{"cited_title":"Vp-sto: Via-point-based stochastic trajectory optimization for reactive robot behavior,","cited_arxiv_id":null,"evidence_quote":"Provides the via-point stochastic trajectory optimization baseline whose offline computation times are compared in Table I."},{"cited_title":"On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,","cited_arxiv_id":null,"evidence_quote":"IPOPT, the interior-point solver used to solve both the offline trajectory optimization (3) and the online replanner's linear constrained quadratic program (13)."},{"cited_title":"Reducing the barrier to entry of complex robotic software: a MoveIt! case study,","cited_arxiv_id":null,"evidence_quote":"Supplies MoveIt, the software environment in which the RRTConnect baseline is run."},{"cited_title":"RRT-connect: An efficient approach to single-query path planning,","cited_arxiv_id":null,"evidence_quote":"Defines the RRTConnect single-query planning algorithm used as the baseline in the real moving-object comparison."},{"cited_title":"Time-optimal trajectory generation for path following with bounded acceleration and velocity,","cited_arxiv_id":null,"evidence_quote":"Provides the time-optimal trajectory generation used to give the RRTConnect path a time parametrization."}],"review_version":1}